Single-Cell Atlas of Driver Mutations in Acute Myeloid Leukemia (AML)

Introduction

Increased complexity of sub-clonal architecture has been associated with poor outcome in AML (Papaemmanuil et al. NEJM 2016). Currently, assessment of intra-tumor genetic heterogeneity is performed with next-generation sequencing (NGS) using bulk tumor samples and relies on the variance of variant allele frequency among the individual mutations. However, this analysis is inherently confounded by the tumor purity and zygosity of the mutations. To overcome these limitations, we recently developed a high-throughput single-cell DNA sequencing platform using droplet microfluidics (Mission Bio Inc.) and showed the feasibility of genotyping primary AML samples at single-cell resolution (Pellegrino et al. Genome Research 2018). Here we used this novel platform in a large cohort of AML samples to characterize the clonal heterogeneity of AML and its evolution at relapse.

Methods

In total, 76 bone marrow (BM) samples from 68 AML patients (pts) were single-cell genotyped using the Mission Bio platform. In order to avoid allelic imbalance, most of the samples (66/76, 87%) were obtained from normal karyotype (NK) AML pts. The platform covered 40 amplicons in 19 recurrently mutated AML genes (median 31x coverage/amplicon/cell [IQR 22-41]). Fastq files were processed using the proprietary pipeline for adapter trimming, sequence alignment, barcode demultiplexing, and genotype and variant calling. Loom files were loaded to Tapestri Insights software for variant filtering. As a reference, all samples were concurrently sequenced by the conventional bulk NGS using targeted capture sequencing (N=64) or whole exome sequencing (N=12). An average allele drop-out (ADO) rate was inferred by the genotype of known single nucleotide polymorphisms that were incorporated into the platform.

Results

In total, 333,731 cells were genotyped from 76 AML samples (median 4,423 cells/sample [IQR 2,801-5,844]). The single-cell DNA sequencing detected 208 driver mutations in 76 samples with median 3 mutations per sample (IQR: 2-3). Most commonly detected mutations were NPM1 (N = 28, 13%), followed by DNMT3A (N = 24, 12%), SRSF2 (N = 24, 12%), FLT3 (N = 22, 11%), and IDH2 (N = 21, 10%), which is in accordance with the genetic landscape for NK AML. All mutations detected by the single-cell sequencing were also confirmed by the bulk NGS. The median ADO rate was 8.5% (IQR 6.8-10.4). We detected median 5 [IQR 4-8] sub-clones per sample by the single-cell sequencing. The platform unambiguously detected co-occurrence and mutual exclusivity among the driver mutations at a single-cell level. For instance, the single-cell sequencing of the samples carrying NRAS/KRAS, double NRAS, double RUNX1, IDH1/IDH2, FLT3-ITD/FLT3-TKD, or NRAS/PTPN11 mutations showed that these two mutations in the same molecular pathway were in different cellular population. In contrast, the platform also detected co-occurrence of multiple mutations in a single-cell. For example, we detected a single cell population with a clear co-occurrence of DNMT3A,FLT3-ITD, and NPM1, the most commonly co-occurring mutations in AML. Computational analysis of the single-cell genotype data by the stochastic search algorithm generated phylogenetic trees of the driver mutations in AML. DNMT3A, IDH1, IDH2, and U2AF1 were frequently detected as a trunk mutation, while mutations in FLT3, NRAS, and NPM1 were frequently detected as branch mutations.

Analysis of 14 baseline and relapse paired samples revealed the remodeling of clonal architecture at relapse in 7 pts. Relapsed samples tended to have simpler clonal architecture with less sub-clones compared to the baseline (7 vs. 4, P = 0.169), suggesting the clonal selection process during the therapy. In 54 pts who were previously untreated and had single-cell genotype information on baseline BM, the pts with ≥ 10 sub-clones had significantly worse overall survival than pts with < 10 sub-clones (2-year survival 17% vs. 43%, P = 0.0468).

Conclusion

The high-throughput single-cell DNA sequencing of 76 AML samples generated an atlas of driver mutations in 333,731 AML cells. The platform uncovered detailed evolutionary history of driver mutations in AML and unambiguously visualized co-occurrence and mutual exclusivity of driver mutations at a single-cell level, features that are not observable with conventional bulk NGS. Our data also suggest the prognostic implication of intra-tumor heterogeneity in AML.